> How could we give LLMs the ability to "pay attention" to different parts of images, as needed, so they can make back-and-forth comparisons between parts of different images to solve these kinds of visual reasoning tasks?

I’ve got good news